Deep Learning Essentials

Pooling or subsampling layer

A pooling or subsampling layer often immediately follows a convolution layer in CNN. Its role is to downsample the output of a convolution layer along both the spatial dimensions of height and width. For example, a 2 x 2 pooling operation on top of 12 feature maps will produce an output tensor of size [16 x 16 x 12] (see the Example of Pooling/Subsampling layer figure).

The primary function of a pooling later is to reduce the number of parameters to be learned by the network. This also has the additional effect of reducing overfitting and thereby increasing the overall performance and accuracy of the network.

There are multiple techniques around pooling. Some of the most common pooling techniques are:

• Max pooling ...